There is a variety of pressure sensors available, however few if none are applicable for use in an artificial skin system. Requirements include robustness regarding wear and tear, easy low-cost mass production and a space-saving, fast read-out mechanism. The skin should be soft and exhibit mechanical properties comparable to that of human skin.
Our aim is to develop robust yet accurate touch sensors for robotic hands. At this, our work is guided by the mechanical and information processing capabilities of the human sensory system. The artificial skin consists of soft, elastic material filled with carbon black particles. We experiment with both rubber and injection-moldable ultrasoft polymers as primary materials. When applying pressure to the composite its resistance decreases, which is due to the creation of conductive pathways from the carbon black. Wires running through the material at different layers are then used to locally measure resistance changes. The first prototype uses two layers, which are orthogonal to each other so as to maximize the number of read-out points.
The resulting system can easily be varied in size, shape and spatial resolution and can be fitted onto 3D surfaces because of its inherent flexbility. By changing the number of wires as well as their location, spatial resolution can be influenced and does not have to stay constant within a single skin patch. While a first application will be single finger tip sensors, covering a whole robotic hand with an artificial skin would be the next stage.
Interpreting artificial skin signals
Initially a simple 2- layer feed forward network was used for interpreting the values from the skin readout system. We used four different objects to apply force to the skin patch with the goal of classifying them correctly. Forces were applied using a spherical, cylindrical, conical and a 4- point indenter. The neural network was trained using two sets for each indenter and then tested with othertwo sets for the same indenters in a sequence (spherical, cylindrical, conical and a 4- point indenter)
The neural network was able to classify the 8 sets correctly as shown in Fig2. The skin was indented two times successively by the same indenter in a sequence specified above. This concluded that the shape (geometry) of the indenter was correctly classified using simple neural networks.
Similar indenting experiments were done also using a real time system. A recurrent neural network was modeled with an objective of classifying different indenting position on the skin patch independent of the type of indenter. Readings were obtained at 5 different positions on the skin patch using three different indenters by applying repetitive force of 5N. We were able to classify the indentations at different positions correctly with a accuracy of about 98%.
Thus our skin patches along with the neural network classified different geometries and positions. Furthermore this approach will be used in future for classifying different forces (magnitude and direction) along with movements.